SpAItial AI introduced Echo-3, "a state-of-the-art world model" that "turns phone captures into photorealistic, metric-scale 3D worlds," from a single room to multi-floor buildings, outdoor and public spaces. The company describes it as its latest physically grounded world model, specialized for reality capture, and says it is available today. A gallery of captures shows the results.
How it works
Cofounder Matthias Niessner's launch thread gives the key idea: treat reconstruction as conditional generation. The model works from a short iPhone video and outputs photorealistic 3D Gaussian splats, point clouds and meshes at metric scale, so real-world dimensions are preserved and measurements can be taken inside the result. Coverage of the launch adds that scans can be restyled.
This differs from the earlier Echo models, which generated explorable 3D worlds from text or an image, starting with the original Echo in December 2025. Echo-3 instead takes a walk through a real space and reconstructs it, using the capture to recover far more of the place's detail.
Access
Echo-3 reaches users through a companion iPhone capture app and an enterprise API for companies integrating reconstruction into their own products and workflows. According to coverage, the app needs iOS 18 or later and is free to download, though processing scans uses credits. LiDAR is optional, with Pro iPhones that have LiDAR getting live coverage guidance and a denser scan preview, and Android is coming soon. SpAItial frames the release around physical AI, which it says is bottlenecked by real-world 3D data.
Other world models launched this week include Google DeepMind's Genie 4, which generates interactive simulations rather than reconstructing real places, and earlier World Labs' Atlas and Skyfall's SKY-001.